Research On Image Super-Resolution Reconstruction Method Based On Deep Learning

Image super-resolution reconstruction technology refers to a special method of converting from low-resolution blurred images to high-resolution images in the same scene.In recent years,due to its extensive practical application value and profound theoretical significance,this technology has been widely used in computer vision,image processing and other fields.Since the use of generative adversarial networks for image super-resolution reconstruction can obtain higher quality and better visual experience,research related to generative adversarial networks has received great attention,but the network has a single scale in model reconstruction.And the shortcomings of insufficient high-frequency information,resulting in strong noise and artifacts in the generated image.In order to solve the above problems,this thesis combines the idea of convolutional neural network to study the SRGAN algorithm,and improves the quality of the reconstructed image by improving its network structure and optimizing the training method.The following is an improvement to the generator model: a parallel residual network structure is proposed,and by adding convolution kernels with different fixed sizes to each residual sub-network structure,the relationship between shallow features and deep features is achieved.Efficient fusion makes full use of images of different scales in each network layer,and solves the problem that the model does not fully apply high-frequency feature information and can only obtain single-scale feature information.In the residual sub-network,the deep residual shrinkage network obtained by improving the basic residual network module is used,and the soft threshold function is added and combined with the attention mechanism,which solves the problem of noise-related features in the model and can enhance feature extraction rate and improve the robustness of the model;the basic idea of adding a multi-scale receptive field module is to cascade multiple small convolution kernels to extract detailed features and reduce computational complexity by increasing the width of the network;upsampling is performed alternately with nearest neighbor interpolation and sub-pixel convolution,which enhances the breadth and depth of information exchange and reduces the number of parameters and time complexity.The following improvements have been made to the discriminator: by introducing a spatial and channel mixed attention mechanism into the basic unit of the network,it can help the network to obtain highfrequency information in the feature map more effectively,thereby enhancing the authenticity recognition rate of the recognized image.Using the method of comparison and research,the test results are compared and explained,which proves that the algorithm in this thesis has better results in both objective and subjective evaluation.The image obtained after reconstruction is excellent,which can solve the problems of single scale and insufficient acquisition of high-frequency information existing in the existing network.